Sciweavers

Free Online Productivity Tools i2Speak i2Symbol i2OCR iTex2Img iWeb2Print iWeb2Shot i2Type iPdf2Split iPdf2Merge i2Bopomofo i2Arabic i2Style i2Image i2PDF iLatex2Rtf Sci2ools

9

COLT
2000
Springer

favoriteEmaildiscussreport

87views Machine Learning» more COLT 2000»

Estimation and Approximation Bounds for Gradient-Based Reinforcement Learning

13 years 8 months ago

Estimation and Approximation Bounds for Gradient-Based Reinforcement Learning

Download www.cs.iastate.edu

We model reinforcement learning as the problem of learning to control a Partially Observable Markov Decision Process ( ¢¡¤£¦¥§ ), and focus on gradient ascent approaches to this problem. In [3] we introduced ¨ ¢¡¤£¦¥§ , an algorithm for estimating the performance gradient of a ©¡¤£¦¥¤ from a single sample path, and we proved that this algorithm almost surely converges to an approximation to the gradient. In this paper, we provide a convergence rate for the estimates produced by ¨ ¢¡¤£¦¥§ , and give an improved bound on the approximation error of these estimates. Both of these bounds are in terms of mixing times of the ©¡¤£¦¥¤ .

Peter L. Bartlett, Jonathan Baxter

Real-time Traffic

COLT 2000 | Gradient Ascent Approaches | Machine Learning | Partially Observable Markov Decision Process | Single Sample Path |

claim paper

Related Content

» GradientBased Learning Updates Improve XCS Performance in Multistep Problems

» The Optimal Reward Baseline for GradientBased Reinforcement Learning

» Variance Reduction Techniques for Gradient Estimates in Reinforcement Learning

» Rates of Convergence of Performance Gradient Estimates Using Function Approximation and Bi...

» PACBayesian Policy Evaluation for Reinforcement Learning

» Approximation and Estimation Bounds for Artificial Neural Networks

» CBR for State Value Function Approximation in Reinforcement Learning

» SampleEfficient Evolutionary Function Approximation for Reinforcement Learning

» Nonparametric Return Distribution Approximation for Reinforcement Learning

Post Info
More Details (n/a)

Added	02 Aug 2010
Updated	02 Aug 2010
Type	Conference
Year	2000
Where	COLT
Authors	Peter L. Bartlett, Jonathan Baxter

Comments (0)